Uploaded by International Research Journal of Engineering and Technology (IRJET)

IRJET- A Survey on Vision based Fall Detection Techniques

International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 06 Issue: 04 | Apr 2019
p-ISSN: 2395-0072
A Survey on Vision based Fall Detection Techniques
Shahna Sherin P1, Anju J2
Student, Dept. of Computer Science and Engineering, LBS Institute of Engineering and Technology for
Women, Kerala, India.
2Professor Anju.J, Dept. of Computer Science and Engineering, LBS Institute of Engineering and Technology for
Women, Kerala, India.
Abstract - Falls can occur anytime ,anywhere and human
body conditions are prone to severe accidents in such
situations. Brownsel et al.[1] analysed how much effect
automatic fall detection systems could have on people’s fear
of falling. The people subject to the experiments had
experienced falls in the previous six months. From the study
it was understood that the people felt more confident and
more safe to walk around after using the fall detectors.
falls can be highly dangerous even leading to
death.Unobserved or unattended falls increase the chance of
casualties.Physically weak or disabled people and aged people
have more chances to fall,faint or injure themselves on the
busy roads ,slippery floors etc.Thus it is very important to
design and develop efficient fall detection systems that would
help to provide quick assistance to the victims .A number of
studies have been made in the area of fall detection and the
fall detection systems are often classified based on the type of
sensors used to record or collect the data.This paper reviews
different fall detection systems based on camera vision that
have been developed over the years to detect human falls.
The main purpose of an automated fall detection system is
to identify that a fall has occured and to provide assistance
in a suitable way .Time is a very crucial thing and if timely
assistance is provided to the fallen people, it would also be
easier to provide fast medical care. Existing health
conditions may worsen when the older people lay down
without being noticed for a long period of time on the roads
or on the floor. If the fall is severe, blood loss for a long time
can even lead to death and it is highly unfortunate that
thousands of people die this way, unattended. It is thus very
important to develop efficient fall detection systems that
would help save lives. Timely detection and timely alerts can
definitely reduce the grave situation of people dying
Key Words: Human Fall,Fall detection, computer vision,
Image Processing, Video Content Analysis etc
Falls are a great cause of fatal injuries, especially for the
people who are old, disabled and weak. According to the
reports of World Health Organization approximately 646000
people die from accidental falls each year globally and most
of these cases are reported from poor countries .Falls are
accountable for most of the deaths due to accidental injuries. It is also understood that falls occur frequently among
people of age group 65 and above. In the case of older people
biological changes in their body make them more weak and
thus more prone to fainting and falling. The effects of such
falls can be highly dangerous and it may even lead to long
term hospitalization or much worse, early death. The critical
phase of the fall is when a person meets the ground or the
lower surface with a shock or an impact. After a fall the
person shouldn’t lie down on the ground for too long. The
fall detectors work towards giving fast detection of the fall
event and giving quick alerts to nearby help centres or
concerned individuals so that the people could be saved at
the earliest. Fear of fall is always associated with the fall.
People who have already experienced such accidental falls
tend to develop negative feelings in their mind and may
refrain from daily life activities due to their fear of falling.
Apart from biological changes in the body that make people
go weak and prone to falls, physical factors are also
responsible for this accidents. Slippery floors and roads
,open drainages or sewage channels, potholes on the roads
etc are a major threat and people irrespective of their age or
© 2019, IRJET
Impact Factor value: 7.211
1.1 Different Fall Detection Systems
Human fall detection has always been an area of constant
research and the number of studies in this area has also
increased drastically over the years. The fall detection
systems are classified according to a variety of criterias. Some
researchers have classified them based on the types of
sensors used to collect fall data such as sensors that are
wearable and that can be embedded somewhere for
monitoring like camera ,pressure sensors etc. Other
classifications are based on detecting the phases after the fall
event and the impact of fall etc. Basically, all fall detection
systems have a similar structure. The data that is needed to
be analysed have to be collected first with the help of various
sensors. Fall detection systems have to efficiently
differentiate between activities such as sitting, walking,
running and falling.
The ambience sensor based fall detection could be implemented only in places equipped with the ambience sensors. This creates the problem of noise often getting mixed
with the sensed data. Wearable device based systems are
found to be used more in case of outdoor fall detection
systems. However when it comes to older people they might
not be that interested in being watched or monitored al-
ISO 9001:2008 Certified Journal
Page 4090
International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 06 Issue: 04 | Apr 2019
p-ISSN: 2395-0072
ways. Apart from that, people may forget to wear such
sensing devices and it also has the added disadvantage of
long term exposure to skin. Vision based fall detection
systems overcome this problem of ambience based and
wearable device based sensors to an extent. It is very
complex and has also various factors that limit its
performance such as effects of lighting on the photos or video
,quality of the camera used for vision, multiple objects that
appear in the background etc.However with the introduction
of latest technologies such as convolutional neural networks,
vision based fall detection techniques have improved a lot.
Neural networks are a good example of advanced machine
learning approaches used for fall detection. Neural networks
simulate the working of our brain and consists of so many
layered networks. The inputs pass through all these layers
and finally the output is obtained. The nodes of these
networks have activation functions such as ReLu (Rectified
Linear Unit) activation function, sigmoid function etc which
are chosen according to the type of the problems. There are
several neural network classifiers that have been developed
over the years and CNNs(Convolutional Neural Networks)
have proven to be efficient for image and video classifications. Instead of doing normal matrix multiplications as it
was in the case of old classifiers, convolutional neural
networks as the name suggests, perform convolutions over
frames which helps in a more detailed classification of the
input images. The number of filters used, play a very
significant role in the performance of neural networks. While
dealing with neural networks, computational complexity,
feasibility of in- corporating so many layers etc have to be
checked. Otherwise it would result in reduced quality of the
Vision based fall detection or sensing systems can be seen
as a category of context aware systems. The context aware
systems as the name suggests involve sensors that record or
sense data from the surroundings by continously monitoring
them. Cameras and sensors embedded on floors or on other
surfaces etc thus come under this category. Comparison of
different fall detection systems is very difficult as they cannot
be easily combined under one single section. Even though
the fall detection based on computer vision is difficult, it has
proven to be one of the best ways to detect falls as cameras
could be placed anywhere indoors or outdoors. Many
researchers have taken in to consideration, the issues of
privacy while designing indoor fall detection systems, making
them user friendly. It is very important to address the privacy
issues as the vision based systems detects falls, after
capturing pictures or videos of the people in focus.This paper
presents a study on different vision based fall detection
Following are some of the studies on vision based fall
detection techniques using deep learning concepts:
Lesya Anishchenko in [3] used deep learning and transfer
learning concepts in the videos obtained from surveillance
cameras. AlexNet was the convolutional neural network
used for classification, classifying fall events from 30 sets of
data records. The method used in here was transfer learning.
The method showed good results in terms of sensitivity values, specificity values etc. Zhou et al.[22] developed a fall
detection system using the concepts of convolutional neural
networks and multi sensor fusion. This system involved a
combination of radar and optical camera. Optical camera
captured the pictures of sequence of human actions. Using
STFT(Short Time Fourier Transform),time frequency(TF)
micromotion characteristics of the radar were obtained. Two
kinds of CNNs, Alex-Net and Single Shot Multi-box Detector(SSD)Net were used for classification and recognition of
the action sequences. Lu et al.[4] developed an intelligent
human fall detection system that could detect falls from
video surveillance. Vibe algorithm was used to detect human
bodies. The gabor features were selected as the observation
feature and SVM was the classifier used. Nadi et al.[5]
developed another system using LDA (Linear Discriminant
Analysis).Aspect ratios and angles were calculated and noise
removal was done to delete shadows. A fall detection method
based on 3D CNN(three dimensional convolutional neural
network) was developed by Lu et al.[6] in which an
automatic feature extractor was trained by kinetic data
alone. By applying three dimensional convolutions over the
frames, motion information was obtained from the videos
apart from the spatial features obtained from 2D images.
Vision based fall detection systems capture images and
videos using a good quality camera and using an efficient fall
detection algorithm they classify the actions or events in the
captured data as ”fall” or ”not fall”. The cameras have to be
attached in appropriate places and the approaches based on
vision are computationally complex, as they need high speed
processors. With the development of technologies like neural
networks, a lot of intelligent fall detection systems have been
made. While designing fall detection systems it is crucial to
reduce the number of false alarms, that is the number of
times the system detect a ’non fall ’ event as a fall event
should be reduced. The number of false positive cases
provided by machine learning methods are very less
compared to many other methods. It is highly complex, but it
is the current trend as it provides much better results. Wang
et al.[2] defined a method which was based on human
characteristic matrix.
The classifier used here to
differentiate between the fall and other activities was
SVM(Support Vector Machine) .The human silhouette was
extracted using background subtraction and human
characteristic matrices were formed based on human
posture informations and Hu-moment invariant. These
features were then used for the training of the classifier.
© 2019, IRJET
Impact Factor value: 7.211
The vision based fall detection techniques can also be
classified based on the type of cameras used for motion or
image capture. Following are some of the studies conducted
ISO 9001:2008 Certified Journal
Page 4091
International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 06 Issue: 04 | Apr 2019
p-ISSN: 2395-0072
with single, multiple and depth cameras: Ko et al.[7]
described a single camera based fall detection method where
both 3D depth tracking and EKF(Extended Kalman Filter)
based approach were combined. This paper described the 2D
image based detection procedure of a person in movement
which involved background subtrac- tion, binarization,
filtering of the noise and finally human detection. EKF was
used to get the depth informations from an image. Using a
set of trained images depth maps were computed with the
help of discriminatively-trained Markov Random Field
(MRF). Bian et al.[8] proposed a fall detection approach
based on a depth camera. This method was not affected by
the changes in illumination of light. Improved Randomized
Decision Tree algorithm was used here to extract 3D joints.
Depth cameras helped to reduce the ambiguity of silhouettes
and depth images were used to apply pose corrections on
the extracted 3D joints of the human body. SVM classifier
was then applied on these extracted 3D joints to confirm
falls. Suneung Kim et al.[9] developed a method based on
Extended Kalman Filter that used both two dimensional and
three dimensional information of a dynamic scene. RGB
image was taken as the input and depth map was created
from that input using learning based methods. Particle
Swarm Optimization was used for tracking humans. Three
dimensional human tracking was done with the help of
Extended Kalman Filter. Detection of fall event was followed
by the ringing of an emergency alarm.
detected using the three dimensional velocities which were
obtained from the 3D trajectory.
2)Spatiotemporal features based approach: Using
spatiotemporal features, shape modeling is used to get
detailed information from the human activities.
Foroughi.et.al[17] has developed a method combining
GMM(Gaussian Mixture Models) and HMM (Hidden Markov
Models) to detect falls. Adaptive GMM method was used to
distuingish humans from other moving images and HMM
was used to recognize the fall events. Chih-Yang Lin et al[13]
described a method to identify human falls through different
shape features. Here again, foreground extraction to obtain
human shapes were done using GMM and Motion History
Image(MHI) was used to analyze the behavior of the fall. An
automated system for fall detection was developed by
Sehairi et al.[14] that does simp-SOBS(Simplified Self
Organized Background Subtraction).The different classifiers
used here were RBF SVM(Radial Basis Function Support
Neighbor)classifier and BPNN(Back Propagation Neural
Network).RBF SVM performed slightly better than KNN as it
is an improved version of KNN and also better than BPNN as
the later works more efficiently with larger amounts of data.
Durga Priya et al.[15] analysed fall detection in the obtained videos by combining motion and shape features.This
method focuses on forming contours of the target human bodies based on morphological skeletons in depth
images.From this contours it then extracts local dynamic
shape and motion features.For improving the accuracy of
the fall detection method various shape and motion features
are combined together.S.M.Naqvi et al.[16] proposed one
class boundary classifier method for fall detection from captured videos.Centroid and orientation of voxel persons were
selected as the features to be extracted and a comparison
was made between 4 different boundary methods which are
one class Support Vector Machine method ,k Nearest
Neighbor method and SCMPM(Single Class Minimax
Probability method)by testing on different datasets.
The fall detection techniques could vary depending upon
the features analysed such as spatiotemporal features, head
position, body posture, body shape change, body state
change etc.
1) Three Dimensional Head Position Analysis Based
Approach: Three dimensional head position analysis
involves the evaluation of head position. This helps in
determining the occurence of large scale movement in the
video sequences. It is dependant on head monitoring.
Hazelhoff et al in [10] designed a fall detection system to
detect falls in an unobserved home situation. The foreground
region extraction was done here by using 2 fixed
uncalibrated cameras. A Gaussian multiframe classifier was
used for the classification of fall events and the rejection of
false positives were done by the head tracking module.
PCA(Principal Component Analysis) was used to find the
direction in which the the body’s main axis lie and also the
differences that occured in the x and y directions. The system
gave a 100% accuracy on un- occluded videos and about
90% accuracy on occluded videos. Rougier and Meunier in
[11] used monocular camera images in their proposed
system. Head 3D Trajectories were the main feature used for
detecting falls in this paper. Here, head tracking was done as
often there would be large amount of head movements
during a fall and the head would be always visible in the
images.3D trajectory was extracted using particle filter and
using this the head was tracked and finally fall was
© 2019, IRJET
Impact Factor value: 7.211
3)Shape and Posture Variation Based Approaches:
Foroughi et al.[12] proposed a method for human fall
detection using variations in the human shape. During
feature extrac- tion, ellipses were approximated around the
human bodies and histograms of both vertical and
horizontal projections were constructed. Head position was
also noted here. Motion classification was done using multi
class SVM and for- ward, backward and sideways falls were
detected. Posture changes were determined by OAA(OneAgainst-All Method) that utilized k binary SVMs to classify k
classes and OAO(One-Against-One) method which used
k(k-1)/2 binary SVMs to identify k classes. Khandoker et
al.[18] tried to find out how wavelet based analysis on
gait variables could be effectively used for dealing with the
balance problems among the elderly. Signal recording was
done to obtain MFC(Minimum Foot Clearance),gait data.
ISO 9001:2008 Certified Journal
Page 4092
International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 06 Issue: 04 | Apr 2019
p-ISSN: 2395-0072
Feature extraction involved analysis of the histogram or the
wavelet analy- sis. SVM was used as the classifier which
helped in the recognition of fall patterns. Miaou et al used
an omni camera or MapCam in [19] to capture pictures and
image processing was done over these images. MapCam has
a 360 degree scene capture capacity and the speciality of
this system is that it
is personalized. Different people
have different body figures and according to this there
would be notable differences in the normal state and fall
states. The personal informations of an individual such as
weight, height etc helped to make the detection much more
easy and to decrease the number of false alerts.
Thome and Miguet in [20] proposed a two layered HHM
model (Hierarchical Hidden Markov Model. The first layer’s
elementary motion pattern helped to detect sudden changes
in the body posture such as falls. Lin and Ling [21] discussed
a fall event detection in a domain that was compressed.
Global motion estimation and local motion clustering
methods were applied here to perform object segmentation
for extracting moving objects. The paper mainly focused on
the extraction of the features such as the fall occurence’s
short time period range, quick change in the centroid of the
falling person, and also the falling human’s vertical
projection histogram.
Table -1:Analysis Of Fall Detection Techniques
Improved fall
detection accuracy by
focusing on specified
angles rather than the
normal ellipse.
invariant and
SVM classifier.
Studied the feasibility
of usage of deep
learning techniques in
fall detection.
Pre-trained CNN
Identified that gabor
feature can effectively
describe human
bodies for fall
Detected falls by
extracting features:
aspect ratios and fall
angles.Moving objects
alone are detected in
the frame by effective
removal of shadows.
Achieved high
accuracy on multiple
cameras fall data set
by using visual guide
© 2019, IRJET
Vibe algorithm
and SVM
Analysis and
Cross Validation
LSTM and 3D
Impact Factor value: 7.211
Successfully detected
human falls in the
outdoor scenarios just
by using single
cameras and also
applying depth informations.
Utilised depth cameras
to avoid the
illumination problems
and overcame the
difficulties of existing
methods in extracting
Detected head
positions and tracks
heads which helps in
obtaining good
robustness and also
helps to remove false
Identified the potential
of three dimensional
tracking based fall
detection with a single
camera.Depth maps
were gener- ated from
the outdoors.
Experimented with
realistic datasets and
achieved excellent
results even with low
images.Identified that
Human shape
deformation is a very
effective tool for
detection of falls.
Combined the motion
information and shape
change information to
detect falls.
Employed cost
effective methods for
detection that could be
applied in real
time.Acceleration and
angular acceleration
calculated to improve
accuracy of fall
Detected silhouettes of
moving persons and
used effective
algorithms to estimate
head positions.
Filtering based
tracking and
depth based fall
de- tection.
Enhanced RDT
algorithm and
SVM classifier.
classifier and
Filtering and
Particle Swarm
Mixture Model
and Pro-crustes
shape analysis.
Multi class SVM
Model and
Motion History
FSM algorithm
and RBFSVM,KNN and
ISO 9001:2008 Certified Journal
Page 4093
International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 06 Issue: 04 | Apr 2019
p-ISSN: 2395-0072
Extracted local
features from contours
of depth images and
obtained high fall
detection accuracy by
fusion of several
extracted features.
Extracted 3D video
features by
constructing a voxel
person and classifiers
based on boundary
methods were
Fall detection
successfully done by
analysing histograms
segmented silhouettes
and noting changes in
head positions
Detected balance
impairments among
old people.Gait variables were used to
extract the features to
be fed in to the
Improved fall
detection accuracy by
incorporating personal
informations like
height,weight etc.
Accurate recognition
of fall pose from sitting
and walking pose and
well capable of
working in unspecified
Fall events identified
and detected by
changes in the human
position,changes in
ratio of vertical
histograms and the
time duration of fall
Informations from
multiple sensors are
fused inorder to
improve the accuracy
of fall detection
© 2019, IRJET
Shape and
analysis and
SVM classifier.
SCMPM method
selected as the
optimal method.
MLP Neural
Minimum Foot
used as the gait
variable and
SVM classifier
The paper presents a survey on various vision based
approaches used in fall detection systems. Vision based fall
detection methods based on different aspects such as the
features used to analyze the fall patterns like spatiotemporal
features, head position, change of body shape and posture
are discussed here. The survey provides insight in to the
wide variety of methods and techniques utilised in the vision
based fall detection systems. It cannot be said that one
particular method is better than the other, each has its own
advantage and disadvantage and the methods are selected
according to their usability and feasibility. But still we
could say that usage of convolutional neural networks could
give us a more accurate fall detection as they outperform
almost all the existing classifiers in image and video
MapCam with
360 degree
scene capturing
ability used for
im- age
detection done
by comparing
with threshold
Hidden Markov
Global Motion
Estimation and
from multiple
sensors are
fused inorder to
improve the
accuracy of fall
Impact Factor value: 7.211
Brownsell S ,Hawley MS .Automatic fall detectors and
the fear of falling. Journal of Telemedicine and
Rui-dong Wang,Yong-liang Zhang ,Ling-ping Dong, Jiawei Lu , Zhi- qin Zhang, Xia He.Fall detection algorithm
for the elderly based on human characteristic matrix
and SVM,2015 15th International Confer- ence on
Control, Automation and Systems (ICCAS).
Lesya Anishchenko ; Machine Learning In Video.
Surveillance For Fall Detection, 2018 Ural Symposium
on Biomedical Engineering, Radioelectronics and
Information Technology (USBEREIT).
Hong Lu,Bohong Yang,Rui Zhao,Pengliang Qu,Wenqiang
Zhang.Intelligent Human Fall Detection for Home
Surveillance. 2014 IEEE 11th Intl Conf on Ubiquitous
Intelligence and Computing and 2014 IEEE 11th Intl
Conf on Autonomic and Trusted Computing and 2014
IEEE 14th Intl Conf on Scalable Computing and
Communications and Its Associated Workshops.
Mai Nadi ,Nashwa El-Bendary,Aboul Ella Hassanien,Taihoon Kim.Falling Detection System Based on Machine
Learning.2015 4th International Conference on
Advanced Information Technology and Sensor
Application (AITS).
Na Lu,Xiaodong Ren,Jinbo Song,Yidan Wu.Visual guided
deep learning scheme for fall detection.2017 13th IEEE
Conference on Automation Science and Engineering
Myeongseob Ko ,Suneung Kim, Kyungchai Lee,Mingi
Kim, Kwangtaek Kim.Single camera based 3D tracking
for outdoor fall detection toward smart healthcare. 2017
2nd International Conference on Bio-engineering for
Smart Technologies (BioSMART).
Zhen-Peng Bian,Junhui Hou,Lap-Pui Chau,Nadia
Magnenat- ThalmannFall.Detection Based on Body Part
Tracking Using a Depth Camera. IEEE Journal of
Biomedical and Health Informatics ( Volume: 19,Issue:
2,March 2015).
Suneung Kim, Myeongseob Ko,Kyungchai Lee,Mingi Kim,
Kwangtaek Kim.3D fall detection for single camera
ISO 9001:2008 Certified Journal
Page 4094
International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 06 Issue: 04 | Apr 2019
p-ISSN: 2395-0072
surveillance systems on the street.2018 IEEE Sensors
Applications Symposium (SAS).
Lykele Hazelhoff,Jungong Han and Peter H.N. de. VideoBased Fall Detection in the Home Using Principal
Component Analysis.10th International Conference,
ACIVS 2008, Juan-les-Pins, France, October 20-24, 2008.
Caroline Rougier, Jean Meunier, Alain St-Arnaud, and
Jacqueline Rousseau. Robust Video Surveillance for Fall
Detection Based on Human Shape Deformation.IEEE
Transactions on circuits and systems for video
technology,vol.21,No.5,May 2011 611.
Homa Foroughi , Alireza Rezvanian ,Amirhossien
Paziraee, Robust Fall Detection Using Human Shape and
Multi-class Support Vector Machine, 2008 Sixth Indian
Conference on Computer Vision, Graphics and Image
Chih-Yang Lin , Shang-Ming Wang , Jia-Wei Hong , Li-Wei
Kang , Chung-Lin Huang,Vision-Based Fall Detection
Through Shape Feature,
2016 IEEE Second
International Conference on Multimedia Big Data.
Kamal Sehairi, Fatima Chouireb, Jean Meunier, Elderly
fall detection system based on multiple shape features
and motion analysis, 2018 International Conference on
Intelligent Systems and Computer Vision (ISCV), Fez,
Morocco, 2018, pp. 1-8.
Durga Priya Kumar , Yixiao Yun , Irene Yu-Hua Gu, Fall
detection in RGB-D videos by combining shape and
motion feature,2016 IEEE International Conference on
Acoustics, Speech and Signal Processing (ICASSP).
M. Yu , S.M. Naqvi , A. Rhuma ,J. Chambers ; One class
boundary method classifiers for application in a videobased fall detection system, IET Computer Vision (
Volume: 6 , Issue: 2 , March 2012 ).
Homa Foroughi,Baharak Shakeri Aski,Hamidreza
Pourreza. Intelligent video surveillance for monitoring
fall detection of elderly in home environments.2008
11th International Conference on Computer and
Information Technology.
Ahsan H. Khandoker ,Daniel Lai , Rezaul K. Begg ,
Marimuthu Palaniswami.A Wavelet-Based Approach for
Screening Falls Risk in the Elderly using Support Vector
Machines.2006 Fourth International Conference on
Intelligent Sensing and Information Processing.
S.-G. Miaou, Pei-Hsu Sung, Chia-Yuan Huang. A
Customized Human Fall Detection System Using OmniCamera Images and Personal Infor- mation.1st
Transdisciplinary Conference on Distributed Diagnosis
and Home Healthcare, 2006. D2H2.
Nicolas Thome, Serge Miguet.A HHMM-Based Approach
for Robust Fall Detection.2006 9th International
Conference on Control, Automa- tion, Robotics and
Chia-Wen Lin , Zhi-Hong Ling. Automatic Fall Incident
in Compressed Video for Intelligent
Homecare.2007 16th International Conference on
Computer Communications and Networks.
Xu Zhou ,Li-Chang Qian ,Peng-Jie You, Ze-Gang Ding , YuQi Han.Fall Detection Using Convolutional Neural
Network With Multi- Sensor Fusion.2018 IEEE
International Conference on Multimedia and Expo
Workshops (ICMEW).
© 2019, IRJET
Impact Factor value: 7.211
ISO 9001:2008 Certified Journal
Page 4095